playbook-review

playbook-review is what makes the consigliere learn. The manager seeds heuristics they already hold (“this owner tends to delay → FUP”), and digest-source proposes new ones it notices across digests. This skill curates that loop; the consigliere role and context-status then apply the heuristics.

Where heuristics live

Scoped to the node they apply to — there is no central playbook:

• contexts/<ctx>/playbook.md / projects/<proj>/playbook.md — about that context or project.
• people/<person>.md — about a person (delay tendency, bus-factor).

The lens reads the playbook of the node in scope plus the relevant people/ file, so context stays small.

Two modes

/octopus:playbook-review            # walk the proposal queue
/octopus:playbook-review --seed     # add a heuristic you already hold

Walk the queue

digest-source appends each proposed heuristic to playbook-inbox.md — the observation, the suggested target node, and the (src: …) evidence. For each one you promote it into the target’s playbook.md / people/ file, edit the wording or retarget, or discard it. Processed proposals leave the inbox.

Seed directly

State a heuristic you already hold; the skill confirms the target node and writes it — no queue round-trip.

Grounding — proposals vs seeds

An agent proposal must carry its (src: …) evidence to be promoted (shown for you to verify, reusing audit-grounding). A seed is trusted — it’s your own knowledge and needs no evidence. Either way, writes stay inside consigliere.workspace.

Where it sits

playbook-review completes the consigliere bundle (4 skills + 1 role): consigliere-bootstrap scaffolds, digest-source captures (and proposes), context-status consults, the consigliere role reads, and this curates the heuristics that make all of it sharper over time.